Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 13:58:46 +01:00
Code Issues Packages Projects Releases Wiki Activity

cmake : fix CMake requirement for CUDA (#7821)

Browse Source
This commit is contained in:
Jared Van Bortel 2024-06-10 18:32:10 -04:00 committed by GitHub
parent fd5ea0f897
commit 864a99e7a0
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 15 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
CMakeLists.txt
View File

@ -402,12 +402,26 @@ if (LLAMA_CUBLAS)
endif() endif()
if (LLAMA_CUDA) if (LLAMA_CUDA)
cmake_minimum_required(VERSION 3.17) cmake_minimum_required(VERSION 3.18) # for CMAKE_CUDA_ARCHITECTURES
find_package(CUDAToolkit) find_package(CUDAToolkit)
if (CUDAToolkit_FOUND) if (CUDAToolkit_FOUND)
message(STATUS "CUDA found") message(STATUS "CUDA found")
if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
# 52 == lowest CUDA 12 standard
# 60 == f16 CUDA intrinsics
# 61 == integer CUDA intrinsics
# 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
if (LLAMA_CUDA_F16 OR LLAMA_CUDA_DMMV_F16)
set(CMAKE_CUDA_ARCHITECTURES "60;61;70") # needed for f16 CUDA intrinsics
else()
set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
#set(CMAKE_CUDA_ARCHITECTURES "OFF") # use this to compile much faster, but only F16 models work
endif()
endif()
message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
enable_language(CUDA) enable_language(CUDA)
set(GGML_HEADERS_CUDA ggml-cuda.h) set(GGML_HEADERS_CUDA ggml-cuda.h)
@ -472,21 +486,6 @@ if (LLAMA_CUDA)
else() else()
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ... set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
endif() endif()
if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
# 52 == lowest CUDA 12 standard
# 60 == f16 CUDA intrinsics
# 61 == integer CUDA intrinsics
# 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
if (LLAMA_CUDA_F16 OR LLAMA_CUDA_DMMV_F16)
set(CMAKE_CUDA_ARCHITECTURES "60;61;70") # needed for f16 CUDA intrinsics
else()
set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
#set(CMAKE_CUDA_ARCHITECTURES "") # use this to compile much faster, but only F16 models work
endif()
endif()
message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
else() else()
message(WARNING "CUDA not found") message(WARNING "CUDA not found")
endif() endif()